Computer Vision
Animals-10 Image Classification
A computer vision project that classifies images into 10 animal categories. I used a MobileNetV2-based model and prepared model exports for different deployment environments.
Statistics, Universitas Airlangga
The Task
The project focuses on supervised image classification: given an input photograph, predict one of the 10 categories in the Animals-10 dataset. The work includes preparing the image input, training the classifier, evaluating it on held-out data, and exporting the resulting model.
Model and Data Preparation
The documented architecture uses MobileNetV2 with an additional convolution and pooling head in a Sequential model. Images are represented as 224 by 224 RGB inputs and converted to floating-point values scaled by 255.
The dataset split is 70% for training, 15% for validation, and 15% for testing. Keeping a separate test split allows the reported test result to be distinguished from training performance.
Recorded Results
The run recorded in the repository completed 8 epochs and reported 98.49% training accuracy and 96.28% test accuracy. These figures describe that documented run and its dataset split; performance on new images can vary.
Model Exports
The repository includes exports for TensorFlow SavedModel, TensorFlow Lite, and TensorFlow.js. The TensorFlow.js export is a GraphModel, and its loader should use tf.loadGraphModel. The TensorFlow Lite labels are supplied alongside the model to map predicted indices to class names.